Who Pays For Interoperability Before The Standard Actually Arrives?

Key Highlights

• A coalition of 19 companies formalized the Open Silicon Photonics for AI Systems effort as an Open Compute Project (OCP) workstream on 13 August 2026, publishing a roughly 300-page architecture blueprint and setting Q4 2026 as the target for submission of first specifications.
• A base specification is only a shared framework; design and product specifications are what would let a second supplier qualify, and OCP has not published a certification date, which means an operator committing to co-packaged optics today is still committing to one vendor's implementation.
• Four separate standardization efforts are now running in parallel, with the OCP silicon photonics workstream joined by the Open CPX MSA and its socketed 6.4T optical engine, the XPO MSA and its liquid-cooled 12.8T module, and a separate OCP workstream on optical transceiver reliability co-chaired by Credo and Oracle.
• HyperFRAME Lens 1H 2026 finds 49% of I&O leaders rate support for and connectivity to other platforms a very significant procurement factor, against 36% for support of AI capabilities and acceleration, with software lock-in at 40% and hardware lock-in at 34%.
• Luma Optics, which delivers pre-deployment optimization, firmware and EEPROM programming, and interoperability validation on modules it also supplies, self-discloses more than 500,000 transceiver units deployed across roughly 30,000 systems with a field failure rate below 0.01%.

The News

On 13 August 2026, an industry coalition formalized its co-packaged optics effort as an official Open Compute Project workstream. The group published an architecture blueprint and named Q4 2026 as the target date for submission of first specifications. The workstream matters because it converts an open-ended ecosystem gap into a dated milestone that infrastructure buyers can plan against. What it does not do is set a date for multi-sourcing, since the certification infrastructure that would let independent suppliers qualify sits two specification layers further out and has no published schedule.OCP co-packaged optics workstream and Q4 target

Analyst Take

During a Schwab Network segment last week ahead of Marvell's earnings print, the question that stayed afterward was not about Marvell. It was about who carries the cost of a data center that has decided not to standardize on a single vendor or photonics spec. Vendors are being rewarded for hedging across incompatible architectures, and the operators run the mixed fabric that results.

The bear case deserves a fair hearing, and it is stronger than the volume of specification activity suggests. Co-packaged optics currently accounts for something in the low single digits of AI data center optical modules, concentrated where the power wall binds hardest. Standards bodies mostly ratify what already shipped, and by the time first specifications are submitted in late 2026, the architectural commitments for 2027 and much of 2028 will be set. On that reading, this is documentation of a niche rather than direction for a market. We think the more useful question is not how long the gap lasts. It is that nobody can tell you.

From our perspective, the current fragmentation across four parallel optics standardization tracks creates a high-risk operational window where infrastructure buyers absorb the friction of uncertified, multi-vendor hardware. Rather than waiting for distant certification milestones, companies such as Luma Optics capitalize on this gap by positioning pre-deployment optimization, EEPROM programming, and layer-one validation as essential operational risk mitigation. By de-risking co-packaged optics and high-density transceiver deployments, where single-module failures can trigger entire switch RMAs or disrupt synchronous AI training jobs, this third-party validation layer provides the crucial bridge between today's architectural hedging and tomorrow's standardized fabrics.

What was Announced

The structurally important disclosure is the specification ladder rather than the blueprint's contents. Under the OCP process, a base specification establishes the shared framework. Design specifications and product specifications follow. Only at that later stage does certification become possible, which is what would allow a second supplier to build a qualified part against a common reference. OCP has published a target for the first rung and nothing for the last one.

That sequencing is the whole story. A base specification is a promissory note the ecosystem writes against itself, redeemable only when certification exists to make it good.

The parallel efforts are approaching the same problem from different angles, which is where the picture gets complicated. The Open CPX MSA is defining a socketed 6.4T optical engine so that near-package and co-packaged designs can be second-sourced the way front-panel pluggables have been for a quarter century. The XPO MSA is specifying a liquid-cooled 12.8T module aimed at front-panel density that current pluggable form factors cannot reach under air cooling. A separate OCP workstream, co-chaired by Credo and Oracle, is working on transceiver reliability itself, seeking to let hosts and transceivers coordinate link-health telemetry and trigger automated responses before a flap disrupts a job.

Four tracks. One stated objective. That is worth a real pause, because it describes an industry where the standardization effort is fragmenting along the same lines as the technology it means to unify. Each group has a different answer about which physical boundary gets frozen first, and they are not obviously converging.

Worth being precise about the timeline, because the shorthand in circulation is wrong in both directions. NVIDIA is shipping co-packaged optics switches now, so the technology is not hypothetical and anyone describing it as years away is mistaken. But CPO displacing pluggables is a different claim entirely. Marvell told investors that pluggable modules remain the primary form factor for scale-out networks and that it does not expect this to change, and management framed the scale-up transition from copper to optics as running for several years with both technologies coexisting. Set against a low single-digit share of deployed modules today, the installed base being characterized and tuned right now is not a legacy problem awaiting retirement. It is the volume base for the rest of the decade.

The second-order effect matters more than the timeline. Co-packaging does not remove optics from the network; it relocates them onto the switch package, which changes what an operator specifies, qualifies and stocks. It also moves the failure economics in an uncomfortable direction. A pluggable that misbehaves is swapped in minutes by a technician at the front panel. An optical engine that misbehaves inside the package is a switch RMA. On that reading, moving optics closer to the silicon appears to raise the return on characterizing a link before it is deployed rather than lowering it.

The scale-up protocol layer above is moving in the same direction, with UALink, ESUN and NVLink Fusion all live after Broadcom left UALink to pursue its own Ethernet variant. Optical interoperability is being standardized underneath a control plane that is actively splintering.

Market Analysis

Marvell's print last week is the cleanest available evidence that the market is already trading on this. The company raised its fiscal 2028 outlook by roughly $1.5 billion to approximately $18 billion, and management cited scale-up optics as a major contributor to that raise. It reaches that revenue by engaging across NPO and CPO and multiple scale-up options rather than betting on one protocol. Marvell is not predicting an architecture. It is selling insurance against the outcome.

Astera Labs is running the same play one layer down. On 21 July 2026 it announced Taurus 3.2T Smart Retimers and Redrivers with footprint compatibility across Ethernet, UALink and ESUN. That design choice only pays if the customer has not decided.

Our own HyperFRAME Research Lens data suggests buyers arrived here first. Interoperability outranked AI acceleration as a procurement factor among the 520 I&O leaders in our 1H 2026 study, with software and hardware lock-in both rating as significant independently. Buyers are pricing architectural commitment as risk, which is a reasonable response to a technology base that has not settled.

Which brings us to the interval, and to why its length matters less than its uncertainty. If certification were known to arrive in eighteen months, the rational move for most operators would be to wait, run conservative configurations and absorb some inefficiency until multi-sourcing became possible. That is not the situation. Four tracks are running, none has published a certification date, and they have not agreed on where the interoperable boundary belongs. Buyers are being asked to deploy through a gap of unknown duration. That is what makes external validation a decision rather than a stopgap.

Somebody absorbs the interoperability cost in that gap. Today that work is manual, and it happens at layer one, where mixed-vendor optics meet varied switch, NIC, firmware, and thermal conditions. Published observations from large-scale training runs have documented transceiver faults degrading throughput across synchronous jobs. A validation layer has formed to absorb this.

Luma Optics - The Optics company you should be tracking

We have recently been briefed extensively by Luma Optics, a private company making great strides in the optics space, on its operational role in AI data center networking. The company functions as part of that validation layer that absorbs manual, layer-one interoperability costs. This layer addresses conditions where mixed-vendor optics encounter varied switch, NIC, firmware, and thermal environments. Such validation is critical because documented transceiver faults can degrade throughput across synchronous jobs during large-scale training runs. To address these challenges, Luma Optics delivers specialized pre-deployment optimization. The company also performs firmware and EEPROM programming for optical hardware. Additionally, Luma Optics provides interoperability validation on the optical modules that it directly supplies. Luma Optics self-discloses a deployed base of more than 500,000 transceiver units. These units are deployed across approximately 30,000 systems. The company reports a field failure rate below 0.01%, though HyperFRAME hasn’t validated these results, so can only interpret these self-disclosed metrics as directional.

Looking Ahead

Two things could affect that validation layer, and neither is the one the layer's participants tend to name. The first is that the work gets standardized and absorbed into the module. The Credo and Oracle reliability workstream is aimed at exactly this, seeded with a specification Credo contributed, and if the model standardizes then characterization migrates from a pre-deployment service into a product feature that ships inside the optics. The second is that fragmentation itself drives large operators to build the capability internally, on the logic that if no standard is coming they cannot wait for one. Stephen Manley, a senior principal network engineer at Oracle Cloud, has presented the operator side of this work publicly. Fragmentation creates the need for external validation and the incentive to in-source it, and we would not assume the first effect dominates. The addressable market for the layer is plausibly smaller than the mixed-optics installed base, and anyone underwriting the category should size it that way.

One observation cuts against treating any of this as a standards story. NVIDIA guided a Q4 FY27 gross margin trough of 71% to 72%, citing memory prices running ahead of prior expectations, while Marvell described customers shifting toward CXL memory expansion partly in response to scarcity. Neither framed that as architectural philosophy. Supply constraint appears to be forcing more variety into AI infrastructure right now than the specification calendar is.

Based on what we are observing, certification is going to keep lagging specification, and the gap between them is where the operational cost lives. Marvell's investor day on 6 October should indicate whether the company intends to size scale-up optics separately from custom silicon, which would give the market its first clean look at what architectural optionality is actually worth in revenue terms. That disclosure would tell buyers more about the economics of hedging than the Q4 submission will.

The specification calendar still deserves more attention than it gets. It will not redirect 2027 deployments, and the bear case is correct about that. Its value sits in the refresh cycle, where it sets the terms on which the next architecture is bought. Today's pluggable ecosystem, moving hundreds of millions of units, began the same way, with a framework that ratified existing practice and later produced a certification regime that routinized multi-sourcing. We are tracking:

• Whether design and product specifications follow on a credible schedule,
• Will the four tracks converge or harden,
• Will any supplier outside the founding coalitions qualifies against them
• Does the validation work that fills the gap stays with specialists, gets absorbed into the modules, or migrate in-house at hyperscalers willing to devote the engineering talent and time to it.